vimarsh/repro-float8-2bits-entropy-coding-enables-data-free-model-compression-artifacts / logbook-files /repro_float8_2bits /outputs /proxy_metrics.csv
| kind,num_weights,available_quantization_levels,observed_unique_values,fixed_uniform_2bit_levels,shannon_bits_per_weight,huffman_bits_per_weight,fixed_2bit_bits_per_weight,data_free,calibration_or_training_used,elapsed_seconds,python,numpy,interpretation | |
| synthetic_proxy_not_llm_evaluation,4000000,35,25,4,1.4081807834156383,1.60053375,2.0,True,False,0.584020190872252,3.13.7,2.5.1,"Demonstrates only that entropy coding can have a low expected rate while retaining more than four reconstruction symbols. It does not implement Float8, ANS, GPU decoding, or measure LLM perplexity/instruction following." | |
Xet Storage Details
- Size:
- 605 Bytes
- Xet hash:
- 8bc2f05e790b6da381eb0020c534b3fbee3f89ae59d444991332781551414f5f
·
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.